1

Distributed Media-Aware Rate Allocationfor Video Multicast over Wireless Networks

Xiaoqing Zhu, Thomas Schierl, Thomas Wiegand, Senior Member, IEEE, and Bernd Girod, Fellow, IEEE

Abstract—A unified optimization framework for rate allocation amongmultiple video multicast sessions sharing a wireless network is presented.Our framework applies to delivery of both scalable and non-scalablevideo streams. In both cases the optimization objective is to minimizethe total video distortion of all peers without incurring excessive networkutilization. Our system model explicitly accounts for heterogeneity inwireless link capacities, traffic contention among neighboring links, aswell as different video rate-distortion (RD) characteristics. The proposeddistributed rate allocation scheme leverages cross-layer informationexchange between the MAC and application layers to achieve fastconvergence at the optimal, media-aware allocation.

Performance of the proposed media-aware rate allocation protocolis compared against a heuristic scheme based on TCP-Friendly RateControl (TFRC). In network simulations of standard-definition (SD)video streaming over single or multiple multicast trees, the proposedscheme consistently achieves higher overall video quality than the TFRC-based heuristics. When delivering scalable streams, the flexibility of per-peer rate adaptation inside each multicast tree yields a further slightimprovement in overall video quality over multicast of non-scalablestreams.

Index Terms—video multicast, wireless 802.11 networks, scalable videocoding (SVC), distributed rate allocation, cross-layer design

I. INTRODUCTION

An efficient way for simultaneously serving the same video contentto a group of interested receivers is via multicast delivery. This canbe achieved either at the network layer via IP multicast, or at theapplication layer by logical relays [1]. With the advance of wirelessnetworking technologies, wireless video multicast offers a viablesolution for many applications, e.g., broadband multimedia contentsharing within a residential community.

Rate allocation for video multicast is necessary for accommodatingheterogeneity in receiver capabilities, while avoiding network con-gestion. In wireless networks, this problem is further complicated bytraffic contention among neighboring links and by heterogeneity inthe link speeds. For instance, in IEEE 802.11 wireless networks withcontention-based Media Access Control (MAC) protocols, packetsof the same size sent over a slow link would occupy the sharedwireless channel for a longer duration than those sent over a fastlink [2]. Moreover, the rate utility varies among video streamscarrying different contents: the impact of the same rate increase on anaction movie sequence may be rather different from that on a head-and-shoulder news clip. Therefore, when multiple video multicastsessions share the same wireless network, it is important for the rateallocation scheme to account for heterogeneity in both wireless linksand video rate-distortion (RD) characteristics.

Our earlier work has studied a distributed media-aware rate alloca-tion scheme for multi-user video streaming over wireless networks [3]

X. Zhu and B. Girod are with the Information Systems Laboratory, StanfordUniversity, Stanford, CA 94305, USA, email: zhuxq@stanfordalumni.org,bgirod@stanford.edu.

T. Schierl is with the Fraunhofer Institute for Telecommunications -Heinrich Hertz Institute, Einsteinufer 37, 10587 Berlin, Germany, email:thomas.schierl@hhi.fraunhofer.de.

T. Wiegand is jointly affiliated with the Image Communications Laboratory,Berlin Institute of Technology, Einsteinufer 17, 10587 Berlin, Germany andthe Fraunhofer Institute for Telecommunications - Heinrich Hertz Institute,Einsteinufer 37, 10587 Berlin, Germany, email: thomas.wiegand@tu-berlin.de.

[4] [5]. The scheme aims at minimizing the total video distortionof all participating streams without incurring excessive networkutilization. The optimal solution is achieved in a distributed manner,by maintaining and exchanging congestion prices locally amongneighboring wireless nodes at the MAC layer while adapting the viderate of each stream at the application layer.

In this paper, we propose to extend the same set of designprinciples to video multicast over wireless networks. We focus onapplication-layer multicast, where pre-encoded video contents can beshared over wireless networks consisting of static nodes. In additionto the challenges faced by unicast streaming, rate allocation schemesfor video multicast need to efficiently accommodate peers experi-encing heterogeneous link qualities. It is also important to obtainperiodic feedback from children peers for constant update of rateallocation decisions, while avoiding explosion of acknowledgementpackets across the mutlicast delivery tree.

We address both considerations in our design of a unified rateallocation optimization framework. In particular, we introduce per-hop acknowledgements between each parent and its children, carryingreports of congestion prices as the main form of feedback. Ourframework also accommodates scenarios of both non-scalable andscalable video multicast. When delivering non-scalable video streams,congestion prices are accumulated recursively along each multicasttree before being sent back to the root peer. When deliveringscalable video streams, congestion prices are accumulated aroundthe neighborhood of each wireless link to facilitate graceful qualityreduction at intermediate nodes inside each multicast tree. In bothscenarios, each sending (root or relay) peer calculates the optimalvideo rate based on both the accumulated congestion price and thevideo RD characteristic. The final allocated video rates hence achievethe minimum total video distortion of all peers in all multicastsessions, without overloading the network.

The rest of the paper is organized as follows. Section II reviewsrelated work in video multicast. Section III presents the wirelessnetwork model and the video rate-distortion model used in ourproblem formulations. Sections IV and V explains how to performrate allocation for multicast delivery of scalable and non-scalablevideo streams, respectively. Section VI compares the performance ofthe proposed media-aware rate allocation protocol against a heuristicscheme base on TCP-Friendly Rate Control (TFRC) in variousnetwork simulation scenarios.

II. RELATED WORK

Regardless of whether video multicast is implemented at thenetwork layer or at the application layer, the system needs to addressthe issue of receiver heterogeneity. For a single multicast session,rate allocation among the receivers is a well-studied optimizationproblem [6] [7]. Given non-scalable compressed media streams,the sender needs to adapt its outgoing video rate to accommodatethe receiver with the lowest bandwidth, as in [8]. Alternatively,one can extend the single-stream adaptation approach to organizereceivers into separate trees according to their bandwidth types, asdescribed in the Destination Set Group (DSG) protocol [9]. Given ascalable representation of the compressed video stream, the approach

2

3

2

5

68

7

9

10

1

4

11

SRC 1

SRC 2

Video Rate Controller

Link State Monitor

Relay

Video Rate Controller

Link State Monitor

Scalable Video Encoder

Fig. 1. Overview of a wireless video multicast system, where twovideo multicast sessions share a wireless network. Each video stream isdelivered to all participating peers over an application-layer multicast tree.Acknowledgment packets are sent on a per-hop basis, from each receiverto its parent peer. Media-aware rate allocation is achieved via interactionsbetween MAC-layer Link State Monitors (LSMs) and application-layerVideo Rate Controllers (VRCs). The modules containing italic fonts areonly needed when delivering scalable video streams.

of receiver-driven layered multicast (RLM) allows each receiver toadjust its subscription to video quality layers based on its availablebandwidth [10]. The layered structure of the media stream is alsowell suited to dynamic optimization of the unequal error protectionstrength among the packets [11] [12] [13].

For rate allocation across different multicast sessions, most solu-tions typically aim at achieving some level of fairness among videosessions and TCP flows [14] [15]. The optimization framework in [16]investigates rate allocation for receivers with general rate utilities overa general network, without catering to the specific characteristics ofvideo streams and wireless networks.

Recent studies have tackled the additional challenges arising frommulticast over wireless networks. In [17] and [18], for instance,the RLM scheme is combined with distributed source coding andforward error correction (FEC) techniques to achieve better error-resiliency. For more robustness against topological changes in awireless network, various multipath routing schemes have beenextended to construction of multiple multicast trees, in order toleverage the structured redundancy in multiple-description codedvideo streams [19] [20] [21]. Studies have also indicated the benefitof cross-layer design for a single wireless video multicast session,for instance via joint optimization of power allocation, multicast treeconstruction, and rate allocation [22] [23].

III. SYSTEM MODEL

Figure 1 presents an overview of a wireless video multicast system.The wireless nodes self-organize into a network and deliver videostreams along application-layer multicast trees. In the following, wefirst introduce the wireless network model and video rate-distortion(RD) model for formulating the multicast rate allocation problem. Wethen describe the key components comprising the proposed media-aware multicast rate allocation protocol.

A. Wireless Network Model

We index each wireless link with l and each video multicast sessionwith s. The set of all links within a wireless network is denoted as L;the set of all multicast sessions, S. The application-layer multicasttree T s for Session s ∈ S consists of all links traversed by that videostream. Link l succeeds Link l′ in multicast tree T s, l′ ⇒ l, if thesource of Link l is the destination of Link l′. In Fig. 1, for instance,Link 5→ 8 succeeds Link 2→ 5 in the first multicast tree.

We assume in this work that transmission errors are effectivelyremedied by error control techniques such as adaptive modulation,channel coding, and retransmissions [24]. Consequently, the wirelesschannel exhibits time-varying throughput over each link, with noresidual packet losses observed at the application layer [25] [26]. 1

We define throughput Cl as the maximum achievable data rate overLink l, when the rest of the network is not transmitting. Total trafficrate over that link is Fl = F ′l +

∑s:l∈T s R

sl , including both the

rate of non-video traffic F ′l and the allocated rates Rsl ’s of all videosessions traversing that link.

In wireless networks with contention-based MAC protocols suchas IEEE 802.11 [27], nearby links l and l′ typically cannot transmitat the same time, and are deemed to interfere with each other.Interference among links is denoted as l ./ l′, and is reciprocal.In other words, l′ ./ l if and only if l ./ l′. For Link li→j , wedefine the set of links whose source or destination node is withintransmission range of either Node i or Node j as its interference setLl. By definition, l′ ./ l,∀l′ ∈ Ll.

We define link utilization as the fraction of time during whicheach link is active: ul = Fl/Cl. It is obvious that the same trafficrate incurs higher utilization over a slower link than over a fasterlink. Consequently, total utilization within each interference set Ll isconstrained by:

ul =∑l′∈Ll

ul′ < γ, (1)

where γ < 1 is an over-provisioning factor. The extra headroomis needed to absorb various effects not included in our model,such as random backoff in a CSMA/CA network to resolve trafficcontention over the shared wireless media, or inaccurate estimates ofinstantaneous link throughput.

B. Video Rate-Distortion Model

The rate-distortion (RD) tradeoff of each video stream is describedusing a parametric model [28]:

Ds(R) =θs

R−Rs0+Ds

0, (2)

where the parameters Ds0, Rs0 and θs can be fitted to empirical RD

data points using nonlinear regression techniques. They are updatedfor every Group Of Pictures (GOP) in the pre-encoded video stream.

For non-scalable video streams, the video rate over each multicasttree is constrained to the rate chosen by the root peer: Rsl = Rs, ∀s ∈S. Rate adaptation is achieved by means of bitstream switchingamong multiple pre-encoded versions of the video sequence.

For scalable video streams, the video rate can be determinedseparately for each peer inside each multicast tree, so long as the rateof a child does not exceed that of its parent: Rsl ≤ Rsl′ for l′ ⇒ l.Rate adaptation can be performed at the root and at each relayingpeer, by means of sending or forwarding subsets of enhancement-layer video packets, a mechanism to be explained in greater detail inSection V.

IV. RATE ALLOCATION FOR NON-SCALABLE STREAMS

A. Optimization Objective

The goal of multicast rate allocation is to maximize the overallviewing experience of all participating peers without incurring exces-sive network utilization. Based on the subjective viewing test results

1While this assumption is valid for wireless networks with static nodesin our setting, it may not hold in networks with mobile nodes and frequenttopological changes. We leave the exploration of mobile networks as futureresearch.

3

described in [29], we choose to minimize the weighted sum of MSEdistortion of all streams:

min∑s∈S w

sNsDs(Rs) (3)

s. t. Rs > Rsmin, ∀s ∈ S (4)

Rs < Rsmax, ∀s ∈ S (5)

ul < γ, ∀l ∈ L. (6)

In (3), the contribution from Stream s is weighted by both the numberof peers in its multicast tree Ns and a user-specified weighting factorws.2 Presence of the user-specified weighting factors allows greaterflexibility our scheme to accommodate different user preferences orpriorities. For instance, a preferred user is able to obtain higherrate and quality for a rather simple video sequence by assigninga higher weighting factor to its multicast stream. In addition, theconstraints (4) and (5) bound the allocated rate within the rangeprovided by available encoded bitstreams. The constraint (6) guardsagainst excessive total utilization within each interference set.

One can easily verify that the optimization in (3) – (6) has a convexobjective function with linear constraints. We only consider the casewhen the problem is feasible, and assume that the case where streamsat minimum rates still violate the total channel utilization constraintis handled separately by some access control mechanism. If all linkstates and all video RD parameters were available at a central entity,the optimal solution could be calculated using standard numericaltechniques such as the interior point method [30]. However, theoverhead in collecting such global information may not scale wellwith growing network size or stream density. In practice, therefore,a distributed solution is preferable.

B. Distributed Solution

We now show how the pricing-based rate control algorithm pro-posed in [31] can be extended solve the multicast rate allocation prob-lem in (3) - (6). The original optimization objective of maximizinggeneric logarithmic utility functions is substituted with minimizingvideo-specific parametric RD functions. In addition, the resourceconstraints now correspond to total channel utilization within eachinterference set, instead of total rate over individual links in a wirednetwork. The algorithm can be decomposed into the following twoiterative steps.

1) Congestion Price Update: A non-negative congestion price λlis associated with each interference set Ll. Its value is updatedperiodically, according to instantaneous residual rate over Link l:

λl(t) = max[λl(t− τ) + κ(ul − γ)Cl, 0]. (7)

In (7), τ indicates the price update interval and κ is a scaling factorcontrolling the update step sizes.3. For λl > 0, the price updateis proportional to instantaneous excess total utilization over Ll andlink throughput Cl. The intuition behind this is that λl increases iftotal channel time utilization ul temporarily exceeds the specifiedlimit γ in order to induce rate reduction by all streams affecting Ll.Conversely, as long as ul is below the target γ, the correspondingcongestion price should keep decreasing to encourage higher ratesfrom all contributing streams.

2Note that the the proposed optimization objective tends to allocate lowerrates for streams traversing multiple congested interferences sets, similar asTCP’s bias against streams experiencing long round-trip-times.

3It will become clear later in (8)-(9) that the unit of the congestion price λshould be MSE distortion divided by rate. Consequently, the unit of κ shouldbe MSE distortion divided by rate squared. From now on, we will denote theunit of λ with MSE/Mbps, and that of κ with MSE/Mbps2.

Rate

Cos

t

Fig. 2. The allocated video rate Rsopt is determined by the accumulatedcongestion price Λs, by the weight of importance ws, and by the videoRD function Ds(R).

2) Video Rate Update: The rate of each multicast tree is updatedas:

Rs =

Rsopt = Rs0 +

√wsNsθs

Λs , Rsmin ≤ Rsopt ≤ RsmaxRsmin, Rsopt < Rsmin

Rsmax, Rsopt > Rsmax

,(8)

where Ds(R) follows the parametric video RD model (2) and

Λs =∑l∈T s

∑l′∈Ll

λl′Cl′

Cl. (9)

The value of Λs summarizes the congestion contribution of allwireless links affected by the multicast tree. In the special case wherethe multicast tree consists of only a single link l, Λs reduces to λl.In general, the contribution from each link in the multicast tree λl isinflated by a factor of 1 +

∑l′ 6=l,l′∈Ll

Cl/Cl′ . Note that the valueof Λs can be calculated in a recursive manner, whereby each nodereports to its parent the accumulated congestion price Λsl of its ownsubtree:

Λsl =

∑

l′∈Llλl′Cl′

Cl+

∑l′:l ⇒ l′ Λsl′ , Ns

l > 0∑l′∈Ll

λl′Cl′

Cl, Ns

l = 0. (10)

Here Nsl indicate the number of children receiving Stream s from

the peer receiving over Link l.As illustrated in Fig. 2, the accumulated congestion price Λs deter-

mines the slope of the linear term counterbalancing the RD tradeoff.Intuitively, a higher video rate is encouraged when the accumulatedcongestion price Λs is lower. When the network becomes congested,increase in congestion prices at bottleneck interference sets leads to ahigher accumulated price Λs and subsequently lower allocated videorates. In addition, the allocated video rate is confined within the rangeof [Rsmin, R

smax], as dictated by (4) - (5).

The iteration between (7) and (8) naturally constitutes a distributedalgorithm: the congestion price update only requires local observationof ul while the video rate update depends only on the end-to-endaccumulated price Λs and the video RD parameters of individualstreams. At each iteration step, the computational complexity forboth congestion price update and video rate update stays constant foreach peer, irrespective of the number of multicast sessions involvedor network size.

At equilibrium, the congestion prices satisfy the following:{λl > 0, ul = γ, orλl = 0, ul < γ.

(11)

In other words, only congestion prices of fully utilized interferencesets have strictly positive values. This corresponds to the Karush-Kuhn-Tucker (KKT) conditions for the constraint (6) in the originaloptimization problem, thereby guaranteeing optimality of the solu-tion [32].

4

C. Rate Allocation Protocol

The proposed media-aware multicast rate allocation protocol com-prises two key components, as shown in Fig. 1: Link State Moni-tors (LSMs) at the MAC layer and Video Rate Controllers (VRCs)at the application layer. The LSM at each wireless node maintainsupdated congestion price information based on local observationsof link throughput and traffic rates. The VRC at the root of eachmulticast tree is in charge of video rate update, whereas VRCs ateach intermediate node mainly generates per-hop acknowledgementpackets and caches accumulated congestion prices for its subtree.Cross-layer information exchange between the LSMs and VRCs isachieved by granting intermediate wireless nodes access to a set ofspecial video packet header fields, as listed in Table I.4

The LSM of each wireless node continuously tracks the throughputand background traffic rates of its outgoing links. Upon relayingof a video packet, it is allowed to read out the advertised videorate Rs from the special header of traversing video packets. Thelocal congestion price λl is periodically updated based on the totalchannel time utilization within local interference sets. The LSM alsoperiodically exchanges link state messages with neighboring nodes,containing entries listed in Table II.

For delivery of non-scalable video streams, the VRC is neededonly at the root of each multicast tree. The convergence criterionfor the rate allocation process is that the calculated rate fluctuationbetween consecutive allocations is smaller than the rate differencebetween adjacent available rate points. At the same time, the receivedaccumulated congestion price should not exhibit an increasing ordecreasing trend. When the allocation has converged to the optimalvalue Rs, the VRC chooses the quality level k of the next GOP suchthat Rsk ≤ Rs < Rsk+1, 1 ≤ k ≤ K. The actual bitstream switchoccurs later, when the first frame in the next GOP is transmitted.Meanwhile, each VRC at an intermediate node keeps a local cacheof congestion prices reported by ACK packet headers from all itschildren. It then reports the sum according to (10) as the accumulatedcongestion price of its subtree in the ACK packet to its parent.

Note from Table I that the size of the video packet header isconstant, and can be easily accommodated by 5 bytes with sufficientaccuracy. On average, the protocol overhead introduced by additionalfields in the video packet header and ACK streams increases on theorder of O(S · P ), where S is the average number of multicastsessions traversing each interference set and P is the average numberof links traversed by each multicast stream in each interference set.While more frequent transmission of ACK packets facilitates fasterconvergence, it also increases the protocol overhead. In this work, wechoose the acknowledgement frequency to be once per video framefor a good balance between protocol overhead and convergence speed.

According to Table II, each entry in the link state message requires6 bytes to ensure sufficient accuracy. Size of the link state messageis approximately proportional to the number of entries in the update,hence proportional to the number of links the node is incident with.On average, the overhead increases on the order of O(N ·L), whereN is average number of nodes within each interference set, and Lis average number of links incident to each node. The frequency ofthe link state message exchanges determines the tradeoff betweenprotocol overhead and its responsiveness. More frequent messageexchanges incur less delay in keeping the link utilization informationconsistent among all neighboring nodes, at the cost of higher channeltime utilization.

4This paper focuses on the conceptual design of the rate allocation protocol.In a real implementation, it is possible to map these special fields as extensionsof existing transport protocol headers, for instance, as supported by RTP [33]header extensions.

Symbol Content Size Range

s Tree/Session ID 1 byte 0 – 255

Rs Advertised Rate 2 bytes 0 – 65535×10−3 Mbps

Λs Congestion Price 2 bytes 0 – 65535×10−2 MSE/Mbps2

TABLE IFIELDS FOR CROSS-LAYER INFORMATION EXCHANGE IN THE VIDEO

PACKET HEADER.

Symbol Content Size Range

l Link ID 1 byte 0 – 255

Cl Link Throughput 2 bytes 0 – 65535×10−3 Mbps

ul Link Utilization 1 byte 0 – 255%

λl Congestion Price 2 bytes 0 – 65535×10−3 MSE/Mbps2

TABLE IIENTRIES IN THE LINK STATE MESSAGE.

V. RATE ALLOCATION FOR SCALABLE STREAMS

In a network containing wireless links with heterogeneous speeds,it is preferable to allow rate adaptation at intermediate relay peersinside each multicast tree. Otherwise the received video qualitieswould be limited by the peer experiencing the slowest link. We adoptthe Scalable Video Coding (SVC) extension of the H.264/MPEG-4AVC standard for this purpose [34] [35].

A. Rate Adaptation with SVC

Figure 3 illustrates the structure of an SVC stream with a GOPlength of 4 frames. The frames labeled T0 form the first temporallayer; the frames labeled T1 and T2 form the second and the thirdtemporal layers respectively. Each frame is encoded into one base-layer picture and one enhancement-layer picture. Note that the base-layer pictures labeled with either T1 or T2 are predicted from theenhancement-layer pictures of neighboring frames. This approach,also known as Medium Granularity Scalability (MGS), provides abalance between high coding efficiency of the base-layer picturesand mismatch error control when the reference enhancement-layerpictures are missing [35].

For a stream with M temporal layers, M + 1 rate points can beattained by sequentially dropping enhancement-layer pictures basedon their temporal priorities. The first rate point corresponds tostreaming base-layer pictures only for all frames; the second ratepoint includes also enhancement-layer pictures belonging to the firsttemporal layer; and so on. The highest rate point includes both base-layer and enhancement-layer pictures for all frames. The RD tradeoffachieved by this approach can be approximated by the parametricmodel (2).

B. Optimization Objective

Similar as in the case of multicast of non-scalable video streams,the goal of the rate allocation scheme is to minimize the totalvideo distortion of all participating peers without incurring excessive

5

Time

T0 T0 T0T1 T1 T2T2T2T2

Base Layer

EnhancementLayer

Fig. 3. Structure of a Group Of Pictures (GOP) in an SVC streamencoded with MGS-SNR and temporal scalability.

network utilization:

min∑s∈S

∑l∈T s w

slD

s(Rsl ) (12)

s. t. Rs > Rsmin, ∀s ∈ S (13)

Rs < Rsmax, ∀s ∈ S (14)

ul < γ, ∀l ∈ L (15)

Rsl ≤ Rsl′ , l′ ⇒ l, ∀l ∈ L, ∀s ∈ S. (16)

In (12), the user-specified weighting factor is denoted by wsl foreach peer in each session, and allows greater flexibility for theschem to accommodate users with different preferences or priorities.The constraints (13) and (14) bound the allocated rate within therange of each scalable video stream. The constraint (15) guardsagainst excessive total utilization within each interference set. Theconstraint (16) states that the allocated rate for a peer cannot exceedthe rate of its parent.

C. Distributed Solution

The optimal distributed solution can, again, achieved by iterativelyupdating local congestion prices while adapting the rate of each peer:

λl(t) = max [λl(t− τ) + κ(ul − γ)Cl, 0] (17)

Rsl =

Rsl,opt = Rs0 +

√wsθs

Λl, Rsl,min ≤ Rsl,opt ≤ Rsl,max

Rsl,min, Rsl,opt < Rsl,min

Rsl,max, Rsl,opt > Rsl,max

,(18)

where the accumulated congestion price Λl accounts for contributionsfor all links within the interference set of Link l:

Λl =

∑l′∈Ll

λl′Cl′

Cl. (19)

In (17), the price update interval is τ and the price update scalingfactor is κ. Following the same intuition as in (7), the congestion priceupdate is proportional to instantaneous excess total utilization overLl and link throughput Cl. A temporarily over congested interferenceset leads to increased value of λl, whereas temporarily under-utilizedinterference set leads decreased value of λl. In (18), a higher videorate is encouraged if the observed accumulated congestion price Λlis low. Conversely, increase in congestion price Λl leads to reducedvideo rate for all streams affecting the interference set Ll. Theallocated rate is further bounded by the minimum and the maximumavailable rates of each stream, and by the rate allocated to the parentpeer: Rsl,min = Rsmin, Rsl,max = min[Rsmax, R

sl′ ], for l′ ⇒ l.

As in the case of non-scalable video multicast, the proposeddistributed algorithm requires simple calculations for congestionprice update and video rate update. At each iteration step, thecomputational complexity at each peer stays constant irrespective ofthe number of participating multicast sessions or the network size.

0 0.5 1 1.5 2 2.5 3 3.5 430

32

34

36

38

40

42

Rate (Mbps)

PSN

R (

dB)

(a) Non-scalable streams

0 0.5 1 1.5 2 2.5 3 3.5 430

32

34

36

38

40

42

Rate (Mbps)

PSN

R (

dB)

CityCrewDowntownHarborIceTerrace

(b) Scalable streams

Fig. 4. Rate-PSNR tradeoff curves of the six test video sequences: City,Crew, Downtown, Harbor, Ice and Terrace. The sequences have spatialresolutions of 704 × 576 pixels and temporal frequencies at 30 framesper second (fps). They are encoded into: (a) non-scalable streams byx264 [36]; (b) scalable streams by SVC JSVM 8.8 [37]. Note that theavailable rates for the non-scalable streams go beyond the range of thegraph shown in (a).

D. Rate Allocation Protocol

For delivery of scalable video streams, the LSMs follow exactlythe same proedures as in the case of multicast of non-scalable videostreams. On the other hand, video rates can be adapted on a per-hopbasis along the multicast distrubtion tree. At the beginning of eachGOP, each VRC informs subsequent children peers of the updatedRD parameters of the video stream via preamble packets. It recordsthe advertised rate of its parent upon receipt of a video packet andextracts the value of accumulated congestion price Λl upon receiptof an ACK packet from one of its children. The optimal rate Rsl isthen calculated according to (18) for each child peer. Same as in non-scalable video multicast, the convergence criterion is that fluctuationin the calculated rate between consecutive allocations is smaller thanthe rate difference between adjacent available rate points withoutexhibiting an increasing or decreasing trend. Upon convergence ofallocation, the VRC then forwards video packets selectively to matchthe allocated rate.

Similar to multicast of non-scalable video streams, overhead ofthe rate allocation scheme introduced by the additional video packetheader scales linearly with the average number of multicast trees andnetwork size, on the order of O(S · P ). The overhead introducedby link state message exchanges scales with network density andnetwork size, on the order of O(N · L).

VI. SIMULATION RESULTS

A. Simulation Setup

We evaluate performance of the proposed media-aware multicastrate allocation protocol in ns-2 [38] simulations over a small wire-less network shown in Fig. 1. The transmission power and receivingthreshold of the nodes are adjusted to achieve a transmission rangeof 55 m. Parameters of SIFS/DIFS/EIFS slot time, random backoffwindow size and retry limits are chosen according to specifications ofthe IEEE 802.11a standard [27]. The basic rate for header and controlpacket transmissions is set to 6 Mbps, whereas the nominal link speedfor payload transmissions varies between 6 and 54 Mbps.5

Six standard-definition (SD) video sequences are considered forstreaming. The sequences have spatial resolutions of 704× 576 pixelsand temporal frequencies of 30 frames per second (fps). For deliveryof non-scalable streams, the video sequences are encoded usingx264 [36], a fast implementation of the H.264/AVC standard [39].The GOP length is 30 frames; the GOP structure is IBBPBBP...,similar to that used in MPEG-2 streams. Different RD tradeoff pointsare obtained by varying the quantization parameter (QP) from 20 to49. For delivery of scalable streams, the test sequences are encoded

5Note that the notion of link speed does not factor in MAC-layer overheadssuch as random backoff and guard time slots between adjacent transmissions,therefore is typically significantly lower than the effective link throughputobserved by the video stream.

6

into scalable video streams using the SVC JSVM 8.8 referencesoftware [37]. Each frame is encoded into one base-layer picturewith QP of 36 and one enhancement-layer picture with QP of 30.The GOP length is 32 frames, corresponding to 6 temporal layersand 7 RD tradeoff points. Figure 4 shows the rate-PSNR tradeoff ofall six sequences for both non-scalable and scalable representations.Note that the adoption of hierarchical B frames in the scalable streamsyields slightly higher coding efficiency than the non-scalable streams.Encoded video frames (or base-layer and enhancement-layer pictures)are further segmented into network packets with maximum size of1500 bytes. Packet transmission intervals are evenly spread out withineach GOP to avoid unnecessary queuing delays due to intra-codedframes.

All simulations presented in this paper adhere to the same set ofprotocol parameter choices: the target utilization is γ = 85%; theprice update scaling factor is κ = 1.0 MSE/Mbps2; the price updateinterval is τ = 10 ms. To balance between fast convergence andprotocol overhead, the video acknowledgment frequency is set as oneACK per frame and the LSM message exchange interval is chosenas 20 ms. The weight of importance is set as unity for all peers in allmulticast trees. The playout deadline for all video streams is chosenas 1 second.

B. TFRC-based Heuristic Scheme

Performance of the media-aware rate allocation protocol is com-pared against a TFRC-based heuristic scheme. We further enhanceTFRC by introducing a virtual random-early-detection (RED) mech-anism at the relay nodes. Such mechanism helps to avoid qualitydegradation caused by actual packet losses due to queue overflow.The relay node monitors its queue size, and calculates the probabilityfor randomly marking a packet as virtual loss according to the sameprinciples in RED queues [40]. Such virtual loss is marked in the the1-bit ECN field in the IP packet header [41]. The receiver calculatesthe average percentage of marked packets as virtual packet loss ratiop and reports this information back to the sender via ACK packets.

For delivery of non-scalable video streams, each relay peer aver-ages the measured round trip times T sl and virtual packet loss ratiospsl reported by all its children, before feeding back such informationto its parent:

psl =

{∑l′:l ⇒ l′ p

sl′ , Ns

l > 0

psl , Nsl = 0,

(20)

T sl =

{∑l′:l ⇒ l′ T

sl′ , Ns

l > 0

T sl , Nsl = 0,

(21)

where Nsl indicate the number of children receiving Stream s from

the peer receiving over Link l. The allocated rate is then calculatedas a function of average end-to-end round-trip time T s and virtualpacket loss ratio ps over the entire multicast tree [42]:

Rs = kB

T s√ps, (22)

with scaling factor k and average packet size B.For delivery of scalable video streams, the allocated rate is cal-

culated on a per-hop basis as a function of the measured round triptime T sl and virtual packet loss ratio psl over the connection betweeneach peer and its parent:

Rsl = kB

T sl√psl. (23)

The rate is further constrained by the rate of the incoming stream,i.e., the rate allocated to the parent peer, to ensure that Rsl ≤ Rsl′for l′ ⇒ l.

0 5 10 15 20 25 30 35 4005

10152025

Thro

ughp

ut

(Mbp

s)

City

0 5 10 15 20 25 30 35 4005

101520

Pric

e (M

SE/M

bps2 )

0 5 10 15 20 25 30 35 40012345

Rat

e (M

bps)

0 5 10 15 20 25 30 35 402025303540

PSN

R (d

B)

0 5 10 15 20 25 30 35 400

50100150200

Time (s)

Del

ay (m

s)

Link 1 2 Link 5 6

Fig. 5. Traces of estimated link capacities, accumulated congestionprices, allocated video rates, corresponding video qualities, and packetdelivery delay. The City sequence streams over the first multicast treeshown in Fig 1. The outgoing link speed of Peer 5 initially starts as54 Mbps, and drops to 12 Mbps between time t = 10s to t = 30s. Allother links operate at 54 Mbps.

C. Allocation for Non-Scalable Streams

1) Single multicast tree: We first consider the simple case of asingle multicast session over the first 7-node multicast tree shown inFig. 1. All links are within the same interference set. Figure 5 showsthe traces of estimated link capacities, accumulated congestion priceΛs over the entire multicast tree, allocated video rates, correspondingvideo qualities in PSNR, and end-to-end packet delivery delayexperienced by Peers 2 and 6. In this experiment, the outgoing linkspeed of Peer 5 initially starts as 54 Mbps, then drops to 12 Mbpsbetween time t = 10s to t = 30s. All other links operate at 54 Mbps.It can be observed that the congestion prices and the allocated videorates quickly reach convergence both before and after the link speeddrop at Peer 5. During the period of low link speed at Peer 5,the allocated rate and resulting video quality of the entire multicasttree are reduced accordingly, as a consequence of the increasedaccumulated congestion price. Since the proposed scheme mindfullyavoids network congestion by limiting the utilization within eachinterference set, it can also leads to low packet delivery delays, wellbelow the latency requirement of 1 second.

Figure 6 compares the media-aware allocation against the TFRC-based heuristic scheme in terms of average video quality, measuredas PSNR of the average video distortion of all peers over the entiremulticast tree. As the outgoing link speed of Peer 5 varies from6 Mbps to 54 Mbps, both schemes achieve similar results in thatthey both try to maximize the allocated rate without overloading thenetwork. Since the TFRC-based heuristic allocation tends to fluctuateover time, it results in a slightly lower average video quality than theproposed media-aware scheme.

2) Multiple multicast trees: Next, we consider the scenario of twovideo sequences streaming over two multicast trees respectively, asdepicted in Fig 1. The outgoing link speed of Peer 5 varies between6 Mbps and 54 Mbps while all other links operate at 54 Mbps.Figure 7 compares the media-aware allocation and the TFRC-basedheuristics scheme for three video sequence pairs. We compare theallocated video rate and quality of each multicast tree, as well asthe average video quality for all peers. Note that the media-awareallocation outperforms the TFRC-based heuristic scheme in terms ofthe overall video quality, measured as PSNR of the average videodistortion of all peers in both trees. The performance gain is morepronounced when there is greater heterogeneity in the wireless link

7

5 10 15 20 25 30 35 40 45 50 5533

34

35

36

37City

Link Speed (Mbps)

Avg

. PSN

R (

dB)

media−awareTFRC

5 10 15 20 25 30 35 40 45 50 5534

35

36

37

38Crew

Link Speed (Mbps)

Avg

. PSN

R (

dB)

media−awareTFRC

5 10 15 20 25 30 35 40 45 50 5541

42

43

44

45Downtown

Link Speed (Mbps)

Avg

. PSN

R (

dB)

media−awareTFRC

5 10 15 20 25 30 35 40 45 50 5529

30

31

32

33

34Harbor

Link Speed (Mbps)

Avg

. PSN

R (

dB)

media−awareTFRC

5 10 15 20 25 30 35 40 45 50 5539

40

41

42

43Ice

Link Speed (Mbps)A

vg. P

SNR

(dB

)

media−awareTFRC

5 10 15 20 25 30 35 40 45 50 5535

36

37

38

39Terrace

Link Speed (Mbps)

Avg

. PSN

R (

dB)

media−awareTFRC

Fig. 6. Average video quality measured as PSNR of the average videodistortion of all peers. One sequence streams over the first multicast treeshown in Fig 1. The outgoing link speed of Peer 5 varies from 6 Mbpsto 54 Mbps while all other links operate at 54 Mbps.

5 10 15 20 25 30 35 40 45 50 550

0.5

1

1.5

2

2.5

3

Link Speed (Mbps)

Allo

cate

d R

ate

(Mbp

s)

City vs. Downtown

Tree 1, media awareTree 1, TFRCTree 2, media awareTree 2, TFRC

5 10 15 20 25 30 35 40 45 50 550

0.5

1

1.5

2

2.5

3

3.5

Link Speed (Mbps)

Allo

cate

d Ra

te (M

bps)

Downtown vs. Crew

5 10 15 20 25 30 35 40 45 50 550

0.5

1

1.5

2

2.5

Link Speed (Mbps)

Allo

cate

d Ra

te (M

bps)

Terrace vs. Ice

5 10 15 20 25 30 35 40 45 50 5528

32

36

40

44

Link Speed (Mbps)

PSN

R (d

B)

Tree 1, media awareTree 1, TFRCTree 2, media awareTree 2, TFRC

5 10 15 20 25 30 35 40 45 50 5526

30

34

38

42

46

Link Speed (Mbps)

PSN

R (d

B)

Tree 1, media awareTree 1, TFRCTree 2, media awareTree 2, TFRC

5 10 15 20 25 30 35 40 45 50 5530

34

38

42

Link Speed (Mbps)

PSN

R (d

B)

5 10 15 20 25 30 35 40 45 50 5530

32

34

36

38

Link Speed (Mbps)

Avg

. PSN

R (d

B)

media awareTFRC

5 10 15 20 25 30 35 40 45 50 5530

32

34

36

38

40

Link Speed (Mbps)

Avg

. PSN

R (d

B)

media awareTFRC

10 20 30 40 5032

34

36

38

40

Link Speed (Mbps)

Avg

. PSN

R (d

B)

media awareTFRC

Fig. 7. Allocated video rate to each multicast tree, average video qualityof each multicast tree, and average video quality measured as PSNR ofthe average video distortion of all peers. Two sequences stream over twomulticast trees as shown in Fig. 1. The outgoing link speed of Peer 5varies from 6 Mbps to 54 Mbps while all other links operate at 54 Mbps.

speeds. When the outgoing link speed of Peer 5 is 6 Mbps, forinstance, the improvement in overall video quality ranges between3.3 dB to 7.2 dB in PSNR.

D. Allocation for Scalable Streams

1) Single multicast tree: When delivering a scalable video streams,rate allocation for peers in the same tree can be different to accommo-date heterogeneity in their link speeds. This is reflected in the traces inFig. 8. In this experiment, the outgoing link speed of Peer 5 initiallystarts as 54 Mbps, then drops to 12 Mbps between time t = 10s tot = 30s. All other links operate at 54 Mbps. It can be noted thatboth congestion prices and allocated video rates rapidly convergencewithin several seconds, both before and after the drop in link speedat Peer 5. During the period when its link speed is at 12 Mbps, thethroughput of Links 5 → 6 and 5 → 8 are reduced accordingly.Consequently, Peer 5 receives higher accumulated congestion pricesfor both its children, hence the allocated rates for Peers 6 and 8 arelower than allocations to other peers in the tree. Similar as in the caseof non-scalable video multicast, the proposed scheme also benefitsfrom proactive avoidance of congestion, hence is able to maintainlow packet delivery delays, well below the latency requirement of 1second.

Figure 9 compares the media-aware allocation against the TFRC-based heuristic scheme in terms of the video quality received byeach peer. As the outgoing link speed of Peer 5 varies from 6 Mbpsto 54 Mbps, the media-aware scheme increases its allocation inaccordance with increase in the wireless link speed, consistently

0 5 10 15 20 25 30 35 4005

10152025

Thro

ughp

ut

(Mbp

s)

Harbor

0 5 10 15 20 25 30 35 400

10203040

Pric

e (M

SE/M

bps2 )

0 5 10 15 20 25 30 35 4001234

Rat

e (M

bps)

0 5 10 15 20 25 30 35 403031323334

PSN

R (d

B)

0 5 10 15 20 25 30 35 400

100200300400

Time (s)

Del

ay (m

s)

Link 1 2 Link 5 6

Peer 2 Peer 6

Fig. 8. Traces of estimated link capacities, accumulated congestionprices, allocated video rates, corresponding video qualities, and packetdelivery delay. The scalably-encoded Harbor sequence streams over thefirst multicast tree shown in Fig 1. The outgoing link speed of Peer 5initially starts as 54 Mbps, and drops to 12 Mbps between time t = 10sto t = 30s. All other links operate at 54 Mbps.

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 2

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 4

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 3

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 5

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 6

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 8

media−aware TFRC

Fig. 9. Video quality of each peer when the scalably-encoded Harborsequence streams over the first multicast tree in Fig 1. The outgoing linkspeed of Peer 5 varies from 6 Mbps to 54 Mbps; all other links operateat 54 Mbps.

allocating higher rates to peers receiving over faster links. In contrast,the TFRC-based heuristics tends to allocate similar rates to all peerswithin the network. When the two outgoing links of Peer 5 operateat the lowest speed, the network only has sufficient resource toaccommodate the Harbor sequence at its lowest rate, therefore resultsfrom both schemes coincide. On the other hand, when all links areoperating at the highest speed, fluctuations in the observed end-to-endround trip times lead to network under-utilization for TFRC-basedheuristics. The media-aware scheme therefore achieves higher overallvideo quality, measured in PSNR of the average video distortion ofall peers. The performance gain ranges between 0.1 dB to 1.9 dBin PSNR for various video sequences and link speeds, as shown inFig. 10. Note also that the video quality from media-aware allocationfor most sequences tends to saturate at the maximum level of thescalable stream, as the outgoing link speed of Peer 5 increases to54 Mbps. The only exception is Harbor, which carries the mostdemanding RD characteristics and a higher maximum rate than whatthe network can support with all links operating at 54 Mbps.

8

5 10 15 20 25 30 35 40 45 50 5532

33

34

35

36

Link Speed (Mbps)

Avg

. PSN

R (

dB)

City

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5534

35

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Crew

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5538

39

40

41

42

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Downtown

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Harbor

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5538

39

40

41

42

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Ice

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5534

35

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Terrace

media−aware TFRC

Fig. 10. Average video quality measured as PSNR of the average videodistortion of all peers. One scalably-encoded video sequence streams overthe first multicast tree shown in Fig 1. The outgoing link speed of Peer 5

varies from 6 Mbps to 54 Mbps, while all other links operate at 54 Mbps.

5 10 15 20 25 30 35 40 45 50 5535

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

City vs. Downtown

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5535

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Downtown vs. Crew

media−aware TFRC

5 10 15 20 25 30 35 40 45 50 5536

37

38

39

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Terrace vs. Ice

media−aware TFRC

Fig. 11. Average video quality measured in PSNR of the average videodistortion of all peers. Two scalably-encoded video sequences stream overtwo multicast trees as shown in Fig. 1. The outgoing link speed of Peer 5

varies from 6 Mbps to 54 Mbps, while all other links operate at 54 Mbps.

2) Multiple multicast trees: Similar observations hold for twovideo sequences over two multicast trees. Figure 11 compares theoverall video quality for all peers in both multicast trees as achievedby the media-aware scheme and by the TFRC-based heuristics. Inaddition to adapting the allocated video rates to their respectiveRD characteristics, the media-aware scheme also allocates highervideo rates for peers receiving over faster links. The TFRC-basedheuristics, in contrast, is oblivious to difference in individual linkspeeds since the observed virtual packet loss ratios and round triptimes are affected by competitions in packet transmissions over allneighboring links. As the outgoing link speed of Peer 5 increasesfrom 6 Mbps to 54 Mbps, the average video quality resulting fromthe media-aware allocation increases accordingly. It outperforms theTFRC-based heuristics by 0.7 - 1.3 dB in PSNR of average videodistortion across all peers.

E. Comparison of Scalable and Non-Scalable Multicast

Finally, we study the effect of scalable representation for videomulticast by comparing the results from the previous two subsections.For completeness, our comparison also include additional results ofnon-scalable multicast of scalable streams. The three alternatives aresummarized as follows:• Scalable Stream, Scalable Multicast (SSSM): Multicast of scal-

able video streams encoded using the SVC JSVM 8.8 referencesoftware [37]. Rate adaptation is performed at each peer insideeach multicast tree, following (18). Simulation results are thesame as those presented in Subsection VI-D.

• Scalable Stream, Non-scalable Multicast (SSNM): Multicast ofscalable video streams encoded using the SVC JSVM 8.8reference software [37]. Rate adaptation is performed at the rootpeer only, following (8) for each entire multicast tree.

• Non-scalable Stream, Non-scalable Multicast (NSNM): Multi-cast of non-scalable video streams encoded using x264 [36],a fast implementation of the H.264/AVC standard [39]. Rateadaptation is performed at the root peer only, following (8) foreach entire multicast tree. Simulation results are the same asthose presented in Subsection VI-C.

Comparisons between SSSM and SSNM highlight the potentialbenefit of scalable multicast from the additional flexibility of rate

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 2

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 4

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 3

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 5

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 6

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5530

31

32

33

34

Link Speed (Mbps)

PSN

R (

dB)

Peer 8

SSSM SSNM NSNM

Fig. 12. Comparison of scalable multicast of scalable streams (SSSM), aswell as non-scalable multicast of both scalable and non-scalable streams(SSNM and NSNM) in terms of video quality of each peer. The Harborsequence streams over the first multicast tree shown in Fig. 1. Theoutgoing link speed of Peer 5 varies from 6 Mbps to 54 Mbps, whileall other links operate at 54 Mbps.

adaptation inside each multicast tree. Performance difference betweenSSNM and NSNM, on the other hand, can be attributed to thedifferent rate-distortion (RD) tradeoffs achieved by the two encodingschemes.

1) Single multicast tree: We first compare the three alternativesin the simple case of streaming the Harbor sequence over a singlemulticast tree. Figure 12 plots the video quality of each peer againstthe link speed of Peer 5. Comparing SSSM against SSNM, it can benoted that scalable multicast tends to improve the video quality ofpeers with higher link speeds, at the expense of lowering the videoquality of peers with lower link speeds. Such difference is moreobvious at intermediate link speeds of Peer 5. Overall, the NSNMscheme achieves similar results as SSNM, with slightly lower videoqualities due to less efficient RD performance of the non-scalablestream.

Figure 13 summarizes the comparison for simulations from allsix sequences. Contrasting SSSM against SSNM, it can be notedthat scalable multicast slightly improves the average video quality.Due to limited rate and quality range of the scalable video streams,the performance gains are only observable when streaming moredemanding sequences Crew and Harbor, or when the link speed ofPeer 5 is at 6 Mbps for City and Terrace. For the same reason, theNSNM scheme benefits from a wider choice of available rates andqualities from multiple non-scalable streams, therefore attains higheraverage video quality when streaming the less demanding Downtownand Ice sequences, or when supporting City and Terrace at higher linkspeeds of Peer 5.

2) Multiple multicast trees: Figure 14 further compares the threeschemes for two video streams over two multicast trees. The benefitof scalable multicast can be observed for two out of three sequencepairs. The performance gain is more obvious at lower link speeds ofPeer 5. As the network becomes more loaded with two streams, thebenefits of both SSSM and SSNM over NSNM start to prevail. Forinstance, the gain in overall video quality is up to 1.2 dB in PSNR, forCity vs. Downtown. In the case of Downtown vs. Crew, on the otherhand, NSNM still achieves a higher average video quality availablefrom a higher-quality version of the non-scalable streams.

9

5 10 15 20 25 30 35 40 45 50 5533

34

35

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

City

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5533

34

35

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Crew

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5540

41

42

43

44

45

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Downtown

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5529

30

31

32

33

34

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Harbor

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5538

39

40

41

42

43

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Ice

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5534

35

36

37

38

39

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Terrace

SSSM SSNM NSNM

Fig. 13. Comparison of scalable multicast of scalable streams (SSSM), aswell as non-scalable multicast of both scalable and non-scalable streams(SSNM and NSNM) in terms of average video quality of all peers.One sequence streams over the first multicast tree shown in Fig. 1. Theoutgoing link speed of Peer 5 varies from 6 Mbps to 54 Mbps, while allother links operate at 54 Mbps.

5 10 15 20 25 30 35 40 45 50 5534

35

36

37

38

Link Speed (Mbps)

Avg

. PSN

R (

dB)

City vs. Downtown

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5536

37

38

39

40

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Downtown vs. Crew

SSSM SSNM NSNM

5 10 15 20 25 30 35 40 45 50 5535

36

37

38

39

Link Speed (Mbps)

Avg

. PSN

R (

dB)

Terrace vs. Ice

SSSM SSNM NSNM

Fig. 14. Comparison of scalable multicast of scalable streams (SSSM), aswell as non-scalable multicast of both scalable and non-scalable streams(SSNM and NSNM) in terms of average video quality of all peers. Twovideo sequences stream over two multicast trees as shown in Fig. 1. Theoutgoing link speed of Peer 5 varies from 6 Mbps to 54 Mbps, while allother links operate at 54 Mbps.

VII. CONCLUSIONS

This paper presents a unified framework for media-aware rateallocation among multiple video multicast sessions sharing awireless network. For delivery of both scalable and non-scalablevideo streams, the proposed scheme aims at minimizing the totalvideo distortion of all peers while limiting network utilization.Our distributed solution leverages cross-layer information exchangebetween MAC-layer link state monitors (LSMs) and application-layer video rate controllers (VRCs), to achieve fast convergenceat the optimal allocated video rates. The resulting allocation isboth media-aware and network-aware. Video streams with moredemanding rate-distortion (RD) characteristics tend to be allocatedhigher rates. In scalable multicast, peers receiving over faster linksget higher rates than peers over slower links.

Performance of the proposed media-aware allocation is comparedagainst a heuristic scheme based on TCP-Friendly Rate Control(TFRC), in simulations of standard-definition (SD) video streamingover single or multiple multicast trees. For delivery of both scalableand non-scalable streams, the media-aware allocation consistentlyoutperforms TFRC-based heuristics in terms of average video qualityof all peers. The additional flexibility of per-peer rate adaptationin scalable multicast yields a further slight improvement in overallvideo quality. On the other hand, the limited rate and quality rangeof a scalable stream tends to limit the improved efficiency ofscalable multicast, when compared with non-scalable multicast ofnon-scalable streams.

One main limitation of the proposed scheme is that it relies oncooperative behavior from all participating multicast sessions toachieve the desired rate allocation results. In reality, however, videostreams sharing a common wireless network may follow differentrate adaptation schemes, for instance based on TCP-friendlyallocation. It therefore remains an intriguing future research topicto augment the design of the proposed media-aware rate allocationscheme for coexistence with other TCP-based flows.

REFERENCES

[1] A. Ganjam and H. Zhang, “Internet multicast video delivery,” Proceed-ings of IEEE, vol. 93, no. 1, pp. 159–170, Jan. 2005.

[2] M. Heusse, F. Rousseau, G. Berger-Sabbatel, and A. Duda, “Performanceanomaly of 802.11b,” in Proc. INFOCOM’03, Twenty-Second AnnualJoint Conference of the IEEE Computer and Communications Societies,vol. 2, San Francisco, CA, USA, Mar. 2003, pp. 836–843.

[3] X. Zhu and B. Girod, “Distributed rate allocation for multi-streamvideo transmission over ad hoc networks,” in Proc. IEEE InternationalConference on Image Processing (ICIP’05), vol. 2, Genoa, Italy, Sept.2005, pp. 157–160.

[4] X. Zhu, P. van Beek, and B. Girod, “Distributed channel time allocationand rate adaptation for multi-user video streaming over wireless homenetworks,” in IEEE International Conference on Image Processing(ICIP’07), San Antonio, TX, USA, Sept. 2007, pp. 69–72.

[5] X. Zhu and B. Girod, “Distributed media-aware rate allocation forwireless video streaming,” IEEE Trans. on Circuits and Systems forVideo Technology, accepted.

[6] T. Jiang, E. W. Zegura, and M. Ammar, “Inter-receiver fair multicastcommunication over the Internet,” in Proc. ACM 9th InternationalWorkshop on Network and Operating Systems Support for Digital Audioand Video (NOSSDAV’99, Basking Ridge, NJ, USA, June 1999, pp. 103–114.

[7] H. Yousefi’zadeh, H. Jafarkhani, and A. Habibi., “Layered media mul-ticast control (LMMC): rate allocation and partitioning.” IEEE/ACMTrans. on Networking, vol. 13, no. 3, pp. 540–553, Mar. 2005.

[8] J.-C. Bolot and T. Turletti, “A rate control mechanism for packetvideo in the Internet,” in Proc. IEEE Annual Joint Conference ofthe IEEE Computer and Communications Societies (INFOCOM’94),Toronto, Canada, June 1994, pp. 1216–1223.

[9] S. Y. Cheung, M. H. Ammar, and X. Li, “On the use of destination setgrouping to improve fairness in multicast video distribution,” in Proc.IEEE Annual Joint Conference of the IEEE Computer and Communica-tions Societies (INFOCOM’96), San Francisco, CA, USA, Mar. 1996,pp. 553–560.

[10] S. McCanne, V. Jacobson, and M. Vetterli, “Receiver-driven layeredmulticast,” in Proc. SIGCOMM’96, vol. 26, no. 4, Stanford, CA, USA,Aug. 1996, pp. 117–130.

[11] X. Li, S. Paul, and M. Ammar, “Layered video multicast with retrans-missions (LVMR): evaluation of hierarchical rate control,” in Proc. IEEEAnnual Joint Conference of the IEEE Computer and CommunicationsSocieties (INFOCOM’98), vol. 3, San Francisco, CA, USA, Mar. 1998,pp. 1062–1072.

[12] P. Chou, A. Wang, and S. Mehrotra, “Error control for receiver-drivenlayered multicast of audio and video,” IEEE Trans. on Multimedia,vol. 3, no. 1, pp. 108–122, Jan. 2001.

[13] W. Tan and A. Zakhor, “Video multicast using layered fec and scalablecompression,” IEEE Trans. on Circuits and Systems for Video Technol-ogy, vol. 11, no. 3, pp. 373–386, Mar. 2001.

[14] X. Li, S. Paul, and M. Ammar, “Multi-session rate control for layeredvideo multicast,” in Proc. SPIE Multimedia Computing and Networking,vol. 3654, San Jose, CA, USA, Jan. 1999, pp. 175–189.

[15] L. Vicisano, L. Rizzo, and J. Crowcroft, “TCP-like congestion control forlayered multicast data transfer,” in Proc. IEEE Annual Joint Conferenceof the IEEE Computer and Communications Societies (INFOCOM’99),New York, NY, USA, Mar. 1999, pp. 996–1003.

[16] K. Kar, S. Sarkar, and L. Tassiulas, “Optimization based rate controlfor multirate multicast sessions,” in Proc. 20th Annual Joint Conferenceof the IEEE Computer and Communications Societies (INFOCOM’01),vol. 1, Anchorage, AK, USA, Apr. 2001, pp. 123–132.

[17] A. Majumdar and K. Ramchandran, “Video multicast over lossy chan-nels based on distributed source coding,” in Proc. IEEE InternationalConference on Image Processing (ICIP’04), vol. 5, Singapore, Oct. 2004,pp. 3093–3096.

[18] J. Wang and K. Ramchandran, “Receiver-driven multicast over wirelesswith distributed source coding and FEC,” in Proc. International Confer-ence on Wireless Communications and Mobile Computing (IWCMC’07),Honolulu, Hawaii, USA, Aug. 2007, pp. 600–605.

[19] S. Mao, X. Cheng, Y. T. Hou, and H. D. Sherali, “Multiple descriptionvideo multicast in wireless ad hoc networks,” in International Confer-ence on Broadband Networks, San Jose, CA, USA, 2004, pp. 671–680.

[20] W. Wei and A. Zakhor, “Multipath unicast and multicast video communi-cation over wireless ad hoc networks,” in Proc. International Conferenceon Broadband Networks (Broadnets’04), San Jose, CA, USA, Oct. 2004,pp. 496–505.

10

[21] ——, “Multiple tree video multicast over wireless ad hoc networks,”IEEE Trans. on Circuits, Systems and Video Technology, vol. 17, no. 1,pp. 2–15, Jan. 2007.

[22] A. Kovacs and I. Godor, “Cross-layer optimized wireless multicast forlayered media,” Computer Networks, vol. 53, no. 7, pp. 1062–1072, May2009.

[23] Y. He, I. Lee, and L. Guan, “Optimized video multicasting over wirelessad hoc networks using distributed algorithm,” IEEE Trans. on Circuits,Systems and Video Technology, vol. 19, no. 6, pp. 796–807, June 2009.

[24] D. Wu, T. Hou, and Y.-Q. Zhang, “Scalable video coding and transportover broadband wireless networks,” Proceedings of the IEEE, SpecialIssue on Multi-Dimensional Broadband Wireless Technologies and Ap-plications, vol. 89, no. 1, pp. 6–20, January 2001.

[25] T. Stockhammer, H. Jenkac, and G. Kuhn, “Streaming video overvariable bit-rate wireless channels,” IEEE Trans. on Multimedia, pp.268–277, April 2004.

[26] Y.Li, A.Markopoulou, N.Bambos, and J.Apostolopoulos, “Joint power-playout control for media streaming over wireless links,” IEEE Trans-actions on Multimedia, vol. 8, no. 4, pp. 830–843, August 2006.

[27] IEEE Standard for Wireless LAN Medium Access Control (MAC) andPhysical Layer (PHY) Specifications, P802.11, Nov. 1997.

[28] K. Stuhlmuller, N. Farber, M. Link, and B. Girod, “Analysis of videotransmission over lossy channels,” IEEE Journal on Selected Areas inCommunications, vol. 18, no. 6, pp. 1012–1032, June 2000.

[29] X. Zhu and B. Girod, “Subjective evaluation of multi-user rate allocationfor streaming heterogeneous video contents over wireless networks,” inProc. IEEE International Conference on Image Processing (ICIP’08),San Diego, CA, USA, Oct. 2008, pp. 3092–3095.

[30] N. Z. Shor, Minimization Methods for Non-differentiable Functions.New York, USA: Springer Series in Computational Mathematics.Springer, 1985.

[31] F. P. Kelly, A. Maulloo, and D. Tan, “Rate control for communicationnetworks: Shadow prices, proportional fairness and stability,” Journal ofOperations Research Society, vol. 49, no. 3, pp. 237–252, Mar. 1998.

[32] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, UnitedKindom: Cambridge University Press, 2004.

[33] H. Schulzrinne, S. Casner, R. Frederick, and V. Jacobson., “RTP: Atransport protocol for real-time applications,” RFC 3550 (Standard), July2003.

[34] ITU-T Recommendation H.264 - ISO/IEC 14496-10, Advanced VideoCoding for Generic Audiovisual services, ITU-T and ISO/IEC JTC 1,2007.

[35] H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the scalable videocoding extension of H.264/AVC,” IEEE Trans. Circuits and Systems forVideo Technology, vol. 17, no. 9, pp. 1103–1120, Sept. 2007.

[36] “x264.” [Online]. Available: http://developers.videolan.org/x264.html[37] Joint Video Team (JVT), “Joint Scalable Video Model (JSVM) reference

software.” [Online]. Available: http://iphome.hhi.de/imagecom\ G1/savce/index.htm

[38] “The Network Simulator NS-2.” [Online]. Available: http://www.isi.edu/nsnam/ns/

[39] ITU-T Recommendation H.264 - ISO/IEC 14496-10, Advanced VideoCoding for Generic Audiovisual services, ITU-T and ISO/IEC JTC 1,2003.

[40] S. Floyd and V. Jacobson, “Random early detection gateways forcongestion avoidance,” IEEE/ACM Transactions on Networking, vol. 1,no. 4, pp. 397–413, Aug. 1993.

[41] K. K. Ramakrishnan, S. Floyd, and D. Black, The Addition of ExplicitCongestion Notification (ECN) to IP, RFC 3168 (Proposed Standard),Sept. 2001.

[42] S. Floyd and K. Fall, “Promoting the use of end-to-end congestioncontrol in the Internet,” IEEE/ACM Trans. on Networking, vol. 7, pp.458–472, Aug. 1999.